Training data is the massive collection of information (books, code, internet forums, academic papers) that an AI model "studies" to learn patterns. This data acts as the model's world-view: its quality, date range, and biases directly dictate how the AI behaves and what it knows.
The internet is the classroom
For an LLM, training data is the "textbook" of human knowledge. The model studies patterns in this data to adjust its internal settings (weights) to get better at guessing what comes next. But here's the reality check: if the training data stops in 2021, the AI has no idea who won the Super Bowl yesterday.
Why training data shapes bias
Because training data reflects the world as it was written about, it inevitably carries our baggage. If the internet over-represents certain cultures or genders in specific roles, the AI will reflect those same stereotypes. This isn't a glitch. It's a mirror. Understanding training data helps you realize why AI might be genius at some topics and totally clueless (or biased) on others.